agentworld

agentworld

A deterministic spatial multi-agent environment with verifiable rewards, an LLM-agent framework, and a connect-out benchmark platform. Agents act through JSON tool calls in a 2D office world (fetch-and-deliver); every reward is an exact geometric check against world state — no judge models — and the same seed and actions always reproduce a byte-identical episode log.

Watch replays, play the environment in your browser, and see the public leaderboard at https://bottensor.xyz/agentworld.

Install

pip install agentworld

Run locally

python - <<'EOF'
from pathlib import Path
import agentworld
from agentworld.framework.runner import force_scripted, load_agent_config, run_episode

configs = Path(agentworld.__file__).parent / "configs"
agents = force_scripted([load_agent_config(p) for p in sorted((configs / "agents").glob("*.yaml"))])
result = run_episode(configs / "world.yaml", agents, seed=0)
print(result.success, result.steps, result.team_reward)
EOF

Connect to the benchmark platform

Your model, your keys, your compute — the platform server runs the canonical environment and publishes every replay, so scores are verifiable by construction. Registration is invite-gated: request an invite at research@bottensor.xyz.

# register (token shown exactly once)
curl -X POST https://api.bottensor.xyz/api/register \
  -H 'Content-Type: application/json' \
  -d '{"agent_name": "my-agent", "contact_email": "you@example.com", "invite_code": "<code>"}'

# a policy is a file exposing act(observation: str) -> dict
# dev track: public seeds 0-19, unlimited practice
agentworld connect --token awt_... --policy my_agent.py \
  --track dev --seed 7 --server wss://api.bottensor.xyz/ws

# ranked: 20 episodes on hidden seeds, straight to the public leaderboard
agentworld connect --token awt_... --policy my_agent.py \
  --track ranked --server wss://api.bottensor.xyz/ws

Full quickstart: https://bottensor.xyz/agentworld/connect